we have months left...

TL;DR

Anthropic's Mythos AI model can autonomously find zero-day exploits at scale, giving the world roughly 6 months to harden defenses before similar capability spreads.

Key Points

• 1.Anthropic's Mythos model autonomously finds and chains zero-day exploits across all operating systems and browsers. It discovered a 27-year-old FreeBSD vulnerability for roughly $50 of compute, shocking early testers in the Glass Wing coalition which includes AWS and Cisco.
• 2.Finding vulnerabilities and patching them are not the same thing. Mythos can dump millions of potential vulnerabilities on engineers' desks, but it cannot autonomously rewrite codebases — humans remain in the loop for all patching, which did not improve in capability.
• 3.Project Glass Wing estimates roughly 6 months to help organizations harden globally before models of Mythos-level capability become more widely accessible or replicated by adversaries.
• 4.Cheap open-source models can already detect the same exploits Mythos found. Testing showed 8 out of 8 small open-weight models identified the flagship FreeBSD exploit, suggesting the capability to 'break stuff' is already broadly available, not locked behind one expensive model.
• 5.Emergence, not intentional design, produced Mythos's cyber capabilities. Anthropic trained for general coding ability; the ability to find and chain exploits arose as an unexpected byproduct, mirroring how OpenAI models unexpectedly solved Erdős math problems.
• 6.XAI's Colossus 2 is training a 10-trillion-parameter model, with a pre-training run estimated at two months, and Meta is expected to have a Mythos-scale model before year-end, meaning the era of powerful models is accelerating, not pausing.
• 7.Viewers are advised to take immediate personal cybersecurity steps including backing up data to an air-gapped hard drive, using a password manager, hardware security keys, encrypted messaging, and reading Karpathy's digital hygiene blog post, regardless of whether the threat materializes.